Chapter 18 parallel databases
1 / 46

Chapter 18: Parallel Databases - PowerPoint PPT Presentation

  • Uploaded on

Chapter 18: Parallel Databases. Introduction I/O Parallelism Interquery Parallelism Intraquery Parallelism Intraoperation Parallelism Interoperation Parallelism Design of Parallel Systems. Chapter 18: Parallel Databases. Parallel machines are becoming quite common and affordable

I am the owner, or an agent authorized to act on behalf of the owner, of the copyrighted work described.
Download Presentation

PowerPoint Slideshow about 'Chapter 18: Parallel Databases' - gisela-chapman

An Image/Link below is provided (as is) to download presentation

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.

- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -
Presentation Transcript

Chapter 18 parallel databases1


I/O Parallelism

Interquery Parallelism

Intraquery Parallelism

Intraoperation Parallelism

Interoperation Parallelism

Design of Parallel Systems

Chapter 18: Parallel Databases


Parallel machines are becoming quite common and affordable

Prices of microprocessors, memory and disks have dropped sharply

Recent desktop computers feature multiple processors and this trend is projected to accelerate

Databases are growing increasingly large

large volumes of transaction data are collected and stored for later analysis.

multimedia objects like images are increasingly stored in databases

Large-scale parallel database systems increasingly used for:

storing large volumes of data

processing time-consuming decision-support queries

providing high throughput for transaction processing


Parallelism in databases

Data can be partitioned across multiple disks for parallel I/O.

Individual relational operations (e.g., sort, join, aggregation) can be executed in parallel

data can be partitioned and each processor can work independently on its own partition.

Queries are expressed in high level language (SQL, translated to relational algebra)

makes parallelization easier.

Different queries can be run in parallel with each other. Concurrency control takes care of conflicts.

Thus, databases naturally lend themselves to parallelism.

Parallelism in Databases

I o parallelism

Reduce the time required to retrieve relations from disk by partitioning

The relations on multiple disks.

Horizontal partitioning – tuples of a relation are divided among many disks such that each tuple resides on one disk.

Partitioning techniques (number of disks = n):


Send the I th tuple inserted in the relation to disk i mod n.

Hash partitioning:

Choose one or more attributes as the partitioning attributes.

Choose hash function h with range 0…n - 1

Let i denote result of hash function h applied to the partitioning attribute value of a tuple. Send tuple to disk i.

I/O Parallelism

I o parallelism cont

Partitioning techniques (cont.): partitioning

Range partitioning:

Choose an attribute as the partitioning attribute.

A partitioning vector [vo, v1, ..., vn-2] is chosen.

Let v be the partitioning attribute value of a tuple. Tuples such that vivi+1 go to disk I + 1. Tuples with v < v0 go to disk 0 and tuples with v vn-2 go to disk n-1.

E.g., with a partitioning vector [5,11], a tuple with partitioning attribute value of 2 will go to disk 0, a tuple with value 8 will go to disk 1, while a tuple with value 20 will go to disk2.

I/O Parallelism (Cont.)

Comparison of partitioning techniques

Evaluate how well partitioning techniques support the following types of data access:

1. Scanning the entire relation.

2. Locating a tuple associatively – point queries.

E.g., r.A = 25.

3. Locating all tuples such that the value of a given attribute lies within a specified range – range queries.

E.g., 10 r.A < 25.

Comparison of Partitioning Techniques

Comparison of partitioning techniques cont

Round robin: following types of data access:


Best suited for sequential scan of entire relation on each query.

All disks have almost an equal number of tuples; retrieval work is thus well balanced between disks.

Range queries are difficult to process

No clustering -- tuples are scattered across all disks

Comparison of Partitioning Techniques (Cont.)

Hash partitioning: following types of data access:

Good for sequential access

Assuming hash function is good, and partitioning attributes form a key, tuples will be equally distributed between disks

Retrieval work is then well balanced between disks.

Good for point queries on partitioning attribute

Can lookup single disk, leaving others available for answering other queries.

Index on partitioning attribute can be local to disk, making lookup and update more efficient

No clustering, so difficult to answer range queries

Comparison of Partitioning Techniques (Cont.)

Comparison of partitioning techniques cont1

Range partitioning: following types of data access:

Provides data clustering by partitioning attribute value.

Good for sequential access

Good for point queries on partitioning attribute: only one disk needs to be accessed.

For range queries on partitioning attribute, one to a few disks may need to be accessed

Remaining disks are available for other queries.

Good if result tuples are from one to a few blocks.

If many blocks are to be fetched, they are still fetched from one to a few disks, and potential parallelism in disk access is wasted

Example of execution skew.

Comparison of Partitioning Techniques (Cont.)

Partitioning a relation across disks

If a relation contains only a few tuples which will fit into a single disk block, then assign the relation to a single disk.

Large relations are preferably partitioned across all the available disks.

If a relation consists of m disk blocks and there are n disks available in the system, then the relation should be allocated min(m,n) disks.

Partitioning a Relation across Disks

Handling of skew

The distribution of tuples to disks may be a single disk block, then assign the relation to a single disk.skewed— that is, some disks have many tuples, while others may have fewer tuples.

Types of skew:

Attribute-value skew.

Some values appear in the partitioning attributes of many tuples; all the tuples with the same value for the partitioning attribute end up in the same partition.

Can occur with range-partitioning and hash-partitioning.

Partition skew.

With range-partitioning, badly chosen partition vector may assign too many tuples to some partitions and too few to others.

Less likely with hash-partitioning if a good hash-function is chosen.

Handling of Skew

Handling skew in range partitioning

To create a a single disk block, then assign the relation to a single disk.balanced partitioning vector (assuming partitioning attribute forms a key of the relation):

Sort the relation on the partitioning attribute.

Construct the partition vector by scanning the relation in sorted order as follows.

After every 1/nth of the relation has been read, the value of the partitioning attribute of the next tuple is added to the partition vector.

n denotes the number of partitions to be constructed.

Duplicate entries or imbalances can result if duplicates are present in partitioning attributes.

Alternative technique based on histograms used in practice

Handling Skew in Range-Partitioning

Handling skew using histograms
Handling Skew using Histograms a single disk block, then assign the relation to a single disk.

  • Balanced partitioning vector can be constructed from histogram in a relatively straightforward fashion

    • Assume uniform distribution within each range of the histogram

  • Histogram can be constructed by scanning relation, or sampling (blocks containing) tuples of the relation

Handling skew using virtual processor partitioning
Handling Skew Using Virtual Processor Partitioning a single disk block, then assign the relation to a single disk.

  • Skew in range partitioning can be handled elegantly using virtual processor partitioning:

    • create a large number of partitions (say 10 to 20 times the number of processors)

    • Assign virtual processors to partitions either in round-robin fashion or based on estimated cost of processing each virtual partition

  • Basic idea:

    • If any normal partition would have been skewed, it is very likely the skew is spread over a number of virtual partitions

    • Skewed virtual partitions get spread across a number of processors, so work gets distributed evenly!

Interquery parallelism

Queries/transactions execute in parallel with one another. a single disk block, then assign the relation to a single disk.

Increases transaction throughput; used primarily to scale up a transaction processing system to support a larger number of transactions per second.

Easiest form of parallelism to support, particularly in a shared-memory parallel database, because even sequential database systems support concurrent processing.

More complicated to implement on shared-disk or shared-nothing architectures

Locking and logging must be coordinated by passing messages between processors.

Data in a local buffer may have been updated at another processor.

Cache-coherency has to be maintained — reads and writes of data in buffer must find latest version of data.

Interquery Parallelism

Cache coherency protocol

Example of a cache coherency protocol for shared disk systems:

Before reading/writing to a page, the page must be locked in shared/exclusive mode.

On locking a page, the page must be read from disk

Before unlocking a page, the page must be written to disk if it was modified.

More complex protocols with fewer disk reads/writes exist.

Cache coherency protocols for shared-nothing systems are similar. Each database page is assigned a home processor. Requests to fetch the page or write it to disk are sent to the home processor.

Cache Coherency Protocol

Intraquery parallelism

Execution of a single query in parallel on multiple processors/disks; important for speeding up long-running queries.

Two complementary forms of intraquery parallelism:

Intraoperation Parallelism – parallelize the execution of each individual operation in the query.

Interoperation Parallelism – execute the different operations in a query expression in parallel.

the first form scales better with increasing parallelism becausethe number of tuples processed by each operation is typically more than the number of operations in a query.

Intraquery Parallelism

Parallel processing of relational operations

Our discussion of parallel algorithms assumes: processors/disks; important for speeding up long-running queries.

read-only queries

shared-nothing architecture

n processors, P0, ..., Pn-1, and n disks D0, ..., Dn-1, where disk Di is associated with processor Pi.

If a processor has multiple disks they can simply simulate a single disk Di.

Shared-nothing architectures can be efficiently simulated on shared-memory and shared-disk systems.

Algorithms for shared-nothing systems can thus be run on shared-memory and shared-disk systems.

However, some optimizations may be possible.

Parallel Processing of Relational Operations

Parallel sort

Range-Partitioning Sort processors/disks; important for speeding up long-running queries.

Choose processors P0, ..., Pm, where mn -1 to do sorting.

Create range-partition vector with m entries, on the sorting attributes

Redistribute the relation using range partitioning

all tuples that lie in the ith range are sent to processor Pi

Pistores the tuples it received temporarily on disk Di.

This step requires I/O and communication overhead.

Each processor Pi sorts its partition of the relation locally.

Each processors executes same operation (sort) in parallel with other processors, without any interaction with the others (data parallelism).

Final merge operation is trivial: range-partitioning ensures that, for 1 j m, the key values in processor Pi are all less than the key values in Pj.

Parallel Sort

Parallel sort cont

Parallel External Sort-Merge processors/disks; important for speeding up long-running queries.

Assume the relation has already been partitioned among disks D0, ..., Dn-1 (in whatever manner).

Each processor Pi locally sorts the data on disk Di.

The sorted runs on each processor are then merged to get the final sorted output.

Parallelize the merging of sorted runs as follows:

The sorted partitions at each processor Piare range-partitioned across the processors P0, ..., Pm-1.

Each processor Pi performs a merge on the streams as they are received, to get a single sorted run.

The sorted runs on processors P0,..., Pm-1 are concatenated to get the final result.

Parallel Sort (Cont.)

Parallel join

The join operation requires pairs of tuples to be tested to see if they satisfy the join condition, and if they do, the pair is added to the join output.

Parallel join algorithms attempt to split the pairs to be tested over several processors. Each processor then computes part of the join locally.

In a final step, the results from each processor can be collected together to produce the final result.

Parallel Join

Partitioned join

For equi-joins and natural joins, it is possible to see if they satisfy the join condition, and if they do, the pair is added to the join output.partition the two input relations across the processors, and compute the join locally at each processor.

Let r and s be the input relations, and we want to compute r r.A=s.Bs.

r and s each are partitioned into n partitions, denoted r0, r1, ..., rn-1 and s0, s1, ..., sn-1.

Can use either range partitioning or hash partitioning.

r and s must be partitioned on their join attributes r.A and s.B), using the same range-partitioning vector or hash function.

Partitions ri and si are sent to processor Pi,

Each processor Pi locally computes riri.A=si.B si. Any of the standard join methods can be used.

Partitioned Join

Partitioned join cont
Partitioned Join (Cont.) see if they satisfy the join condition, and if they do, the pair is added to the join output.

Fragment and replicate join

Partitioning not possible for some join conditions see if they satisfy the join condition, and if they do, the pair is added to the join output.

E.g., non-equijoin conditions, such as r.A > s.B.

For joins were partitioning is not applicable, parallelization can be accomplished by fragment and replicate technique

Depicted on next slide

Special case – asymmetric fragment-and-replicate:

One of the relations, say r, is partitioned; any partitioning technique can be used.

The other relation, s, is replicated across all the processors.

Processor Pithen locally computes the join of riwith all of s using any join technique.

Fragment-and-Replicate Join

Depiction of fragment and replicate joins
Depiction of Fragment-and-Replicate Joins see if they satisfy the join condition, and if they do, the pair is added to the join output.

Fragment and replicate join cont

General case: reduces the sizes of the relations at each processor.

r is partitioned into n partitions,r0, r1, ..., r n-1;s is partitioned into m partitions, s0, s1, ..., sm-1.

Any partitioning technique may be used.

There must be at least m * n processors.

Label the processors as

P0,0, P0,1, ..., P0,m-1, P1,0, ..., Pn-1m-1.

Pi,j computes the join of ri with sj. In order to do so, riis replicated to Pi,0, Pi,1, ..., Pi,m-1, while si is replicated to P0,i, P1,i, ..., Pn-1,i

Any join technique can be used at each processor Pi,j.

Fragment-and-Replicate Join (Cont.)

Fragment and replicate join cont1

Both versions of fragment-and-replicate work with any join condition, since every tuple in r can be tested with every tuple in s.

Usually has a higher cost than partitioning, since one of the relations (for asymmetric fragment-and-replicate) or both relations (for general fragment-and-replicate) have to be replicated.

Sometimes asymmetric fragment-and-replicate is preferable even though partitioning could be used.

E.g., say s is small and r is large, and already partitioned. It may be cheaper to replicate s across all processors, rather than repartition r and s on the join attributes.

Fragment-and-Replicate Join (Cont.)

Partitioned parallel hash join

Parallelizing partitioned hash join: condition, since every tuple in

Assume s is smaller than r and therefore s is chosen as the build relation.

A hash function h1 takes the join attribute value of each tuple in s and maps this tuple to one of the n processors.

Each processor Pireads the tuples of s that are on its disk Di, and sends each tuple to the appropriate processor based on hash function h1. Let si denote the tuples of relation s that are sent to processor Pi.

As tuples of relation s are received at the destination processors, they are partitioned further using another hash function, h2, which is used to compute the hash-join locally. (Cont.)

Partitioned Parallel Hash-Join

Partitioned parallel hash join cont

Once the tuples of condition, since every tuple ins have been distributed, the larger relation r is redistributed across the m processors using the hash function h1

Let ri denote the tuples of relation r that are sent to processor Pi.

As the r tuples are received at the destination processors, they are repartitioned using the function h2

(just as the probe relation is partitioned in the sequential hash-join algorithm).

Each processor Pi executes the build and probe phases of the hash-join algorithm on the local partitions riand s of r and s to produce a partition of the final result of the hash-join.

Note: Hash-join optimizations can be applied to the parallel case

e.g., the hybrid hash-join algorithm can be used to cache some of the incoming tuples in memory and avoid the cost of writing them and reading them back in.

Partitioned Parallel Hash-Join (Cont.)

Parallel nested loop join

Assume that condition, since every tuple in

relation s is much smaller than relation r and that r is stored by partitioning.

there is an index on a join attribute of relation r at each of the partitions of relation r.

Use asymmetric fragment-and-replicate, with relation s being replicated, and using the existing partitioning of relation r.

Each processor Pjwhere a partition of relation s is stored reads the tuples of relation s stored in Dj, and replicates the tuples to every other processor Pi.

At the end of this phase, relation s is replicated at all sites that store tuples of relation r.

Each processor Pi performs an indexed nested-loop join of relation s with the ith partition of relation r.

Parallel Nested-Loop Join

Other relational operations
Other Relational Operations condition, since every tuple in

Selection (r)

  • If  is of the form ai = v, where ai is an attribute and v a value.

    • If r is partitioned on ai the selection is performed at a single processor.

  • If  is of the form l <= ai <= u (i.e.,  is a range selection) and the relation has been range-partitioned on ai

    • Selection is performed at each processor whose partition overlaps with the specified range of values.

  • In all other cases: the selection is performed in parallel at all the processors.

Other relational operations cont
Other Relational Operations (Cont.) condition, since every tuple in

  • Duplicate elimination

    • Perform by using either of the parallel sort techniques

      • eliminate duplicates as soon as they are found during sorting.

    • Can also partition the tuples (using either range- or hash- partitioning) and perform duplicate elimination locally at each processor.

  • Projection

    • Projection without duplicate elimination can be performed as tuples are read in from disk in parallel.

    • If duplicate elimination is required, any of the above duplicate elimination techniques can be used.

Grouping aggregation
Grouping/Aggregation condition, since every tuple in

  • Partition the relation on the grouping attributes and then compute the aggregate values locally at each processor.

  • Can reduce cost of transferring tuples during partitioning by partly computing aggregate values before partitioning.

  • Consider the sum aggregation operation:

    • Perform aggregation operation at each processor Pi on those tuples stored on disk Di

      • results in tuples with partial sums at each processor.

    • Result of the local aggregation is partitioned on the grouping attributes, and the aggregation performed again at each processor Pi to get the final result.

  • Fewer tuples need to be sent to other processors during partitioning.

Cost of parallel evaluation of operations
Cost of Parallel Evaluation of Operations condition, since every tuple in

  • If there is no skew in the partitioning, and there is no overhead due to the parallel evaluation, expected speed-up will be 1/n

  • If skew and overheads are also to be taken into account, the time taken by a parallel operation can be estimated as

    Tpart + Tasm + max (T0, T1, …, Tn-1)

    • Tpart is the time for partitioning the relations

    • Tasm is the time for assembling the results

    • Ti is the time taken for the operation at processor Pi

      • this needs to be estimated taking into account the skew, and the time wasted in contentions.

Interoperator parallelism
Interoperator Parallelism condition, since every tuple in

  • Pipelined parallelism

    • Consider a join of four relations

      • r1 r2 r3 r4

    • Set up a pipeline that computes the three joins in parallel

      • Let P1 be assigned the computation of temp1 = r1 r2

      • And P2 be assigned the computation of temp2 = temp1 r3

      • And P3 be assigned the computation of temp2 r4

    • Each of these operations can execute in parallel, sending result tuples it computes to the next operation even as it is computing further results

      • Provided a pipelineable join evaluation algorithm (e.g., indexed nested loops join) is used

Factors limiting utility of pipeline parallelism
Factors Limiting Utility of Pipeline Parallelism condition, since every tuple in

  • Pipeline parallelism is useful since it avoids writing intermediate results to disk

  • Useful with small number of processors, but does not scale up well with more processors. One reason is that pipeline chains do not attain sufficient length.

  • Cannot pipeline operators which do not produce output until all inputs have been accessed (e.g., aggregate and sort) 

  • Little speedup is obtained for the frequent cases of skew in which one operator's execution cost is much higher than the others.

Independent parallelism
Independent Parallelism condition, since every tuple in

  • Independent parallelism

    • Consider a join of four relations

      r1 r2 r3 r4

      • Let P1 be assigned the computation of temp1 = r1 r2

      • And P2 be assigned the computation of temp2 = r3 r4

      • And P3 be assigned the computation of temp1 temp2

      • P1 and P2 can work independently in parallel

      • P3 has to wait for input from P1 and P2

        • Can pipeline output of P1 and P2 to P3, combining independent parallelism and pipelined parallelism

    • Does not provide a high degree of parallelism

      • useful with a lower degree of parallelism.

      • less useful in a highly parallel system.

Query optimization
Query Optimization condition, since every tuple in

  • Query optimization in parallel databases is significantly more complex than query optimization in sequential databases.

  • Cost models are more complicated, since we must take into account partitioning costs and issues such as skew and resource contention.

  • When scheduling execution tree in parallel system, must decide:

    • How to parallelize each operation and how many processors to use for it.

    • What operations to pipeline, what operations to execute independently in parallel, and what operations to execute sequentially, one after the other.

  • Determining the amount of resources to allocate for each operation is a problem.

    • E.g., allocating more processors than optimal can result in high communication overhead.

  • Long pipelines should be avoided as the final operation may wait a lot for inputs, while holding precious resources

Query optimization cont
Query Optimization (Cont.) condition, since every tuple in

  • The number of parallel evaluation plans from which to choose from is much larger than the number of sequential evaluation plans.

    • Therefore heuristics are needed while optimization

  • Two alternative heuristics for choosing parallel plans:

    • No pipelining and inter-operation pipelining; just parallelize every operation across all processors.

      • Finding best plan is now much easier --- use standard optimization technique, but with new cost model

      • Volcano parallel database popularize the exchange-operator model

        • exchange operator is introduced into query plans to partition and distribute tuples

        • each operation works independently on local data on each processor, in parallel with other copies of the operation

    • First choose most efficient sequential plan and then choose how best toparallelize the operations in that plan.

      • Can explore pipelined parallelism as an option

  • Choosing a good physical organization (partitioning technique) is important to speed up queries.

Design of parallel systems
Design of Parallel Systems condition, since every tuple in

Some issues in the design of parallel systems:

  • Parallel loading of data from external sources is needed in order to handle large volumes of incoming data.

  • Resilience to failure of some processors or disks.

    • Probability of some disk or processor failing is higher in a parallel system.

    • Operation (perhaps with degraded performance) should be possible in spite of failure.

    • Redundancy achieved by storing extra copy of every data item at another processor.

Design of parallel systems cont
Design of Parallel Systems (Cont.) condition, since every tuple in

  • On-line reorganization of data and schema changes must be supported.

    • For example, index construction on terabyte databases can take hours or days even on a parallel system.

      • Need to allow other processing (insertions/deletions/updates) to be performed on relation even as index is being constructed.

    • Basic idea: index construction tracks changes and “catches up”on changes at the end.

  • Also need support for on-line repartitioning and schema changes (executed concurrently with other processing).

End of chapter

End of Chapter condition, since every tuple in

Figure 18 01
Figure 18.01 condition, since every tuple in

Figure 18 02
Figure 18.02 condition, since every tuple in

Figure 18 03
Figure 18.03 condition, since every tuple in